Bacterial ring rot (BRR) in potato (Solanum tuberosum L.) is caused by the gram-positive bacterium Clavibacter michiganensis subsp. sepedonicus (Cms) (Spieck. & Kott.) Skapt & Burkh. This pathogen is regulated by many countries and they maintain a zero tolerance of Cms in certified seed potatoes. The Cms bacteria are mechanically transmitted through routine handling procedures utilized in planting, harvesting, and seed cutting. Exclusion of Cms is enforced through pre- and postharvest visual inspections of seed potatoes and by immunological or nucleic acid diagnostic assays (De Boer et al. 1988; Pastrik 2000). However, the epidemiology of the pathogen remains poorly understood, especially in regards to infection sources and many of the environmental factors influencing symptom expression. This limits the effectiveness of existing detection and control strategies.

Symptoms of BRR are influenced by many factors including environment and genotype (Bonde and Covell 1950; Kawchuk et al. 1997; Logsdon 1967; Nelson and Kozub 1987; Paquin and Genereux 1976; Sherf 1944; Westra et al. 1994; Westra and Slack 1994). Typical symptoms include interveinal chlorosis and stunting followed by wilting, necrosis of the foliage and breakdown of the vascular ring that extends throughout the tuber (De Boer and Slack 1984). Sources of Cms include alternative hosts such as sugar beets or weeds, infected seed potatoes, and facilities or equipment contaminated with Cms that remain a source often for several years probably through the formation of biofilms (Bugbee et al. 1987; Danhorn and Fuqua 2007; List and Kreutzer 1942; Nelson 1980; van der Wolf et al. 2005). Since the incidence of disease is usually low, because of substantial efforts by seed potato certification programs and producer management strategies, more information regarding the expected levels of symptom expression in seed potatoes in any given year would be beneficial in determining the level of inspection and sampling that may be required for detection. However, previous studies examined only a few factors at any given time and little progress has been made in providing a predictive model of symptom expression based on environmental conditions influencing the epidemiology of BRR (Bishop and Slack 1987; Westra et al. 1994; Westra and Slack 1994).

Determining the many environmental interrelationships that may potentially influence the prevalence and intensity of BRR symptoms would require the accumulation and analysis of considerable data of the various factors contributing to the disease. Neural network (NN) models are used to make predictions for complex, non-linear systems with many co-related variables. Agricultural systems usually involve many variables including weather that makes NN models particularly applicable for predicting agricultural outcomes. We have previously used NN modeling to predict maturity of spring wheat in western Canada (Hill et al. 2002), cuticle cracking in greenhouse peppers and tomatoes (Ehret et al. 2008), harvest dates of greenhouse-grown sweet peppers (Lin and Hill 2007), and the weekly yields of sweet peppers grown in commercial greenhouses (Lin and Hill 2008). The objective of this study was to determine if NN modeling could identify environmental factors influencing BRR symptoms to advance our understanding of Cms pathogen-host-environment interactions, predict disease severity, and improve early detection of the pathogen.

Materials and Methods

Pathogen and Inoculum Preparation

An isolate of Cms (LA8503) inciting disease in potato was reisolated each year from field propagated potato tubers, maintained in potato plants in a greenhouse during the winter, and reisolated each spring on Burkholder agar at 23 C. No changes were observed in the pathogen or pathogenecity. Inoculum was rinsed from the agar and Cms diluted to approximately1010 cfu/ml viable cells with sterile H2O immediately prior to inoculation and planting. The level of inoculum was selected to ensure consistent infection of each inoculated tuber between replicates (SE = 0 to 0.8) and provide disease symptoms ranging from no visible symptoms in Teton to advanced wilt and chlorosis of upper leaves in Norland (Kawchuk et al. 1998; Westra et al. 1994).

Potato Inoculation and Propagation

Pre-Elite virus-free seed potatoes of 154 cultivated potato genotypes stored a minimum of 180 days at 4 C were inoculated with Cms isolate LA8503 as previously described (Inglis et al. 1999; Kawchuk et al. 1998). Uniform cut seed pieces of approximately 20 g with at least two sprouts were inoculated using knives dipped into a Cms cell suspension and the bacteria introduced into multiple wounds surrounding each sprout. Inoculation mimics the wounding or cutting of seed that can cause infection. Control seed was treated in a similar treatment but knives were dipped into sterile H2O containing a 1:1000 dilution of quaternary ammonium disinfectant HY-X (Niagra Chemical, Burlington, ON). Genotypes with disease symptom expression ranging from the asymptomatic Teton to the strong symptom expressing Norland were examined over 15 years between 1986 and 2009 in an isolated field plot at Stavely, AB (latitude 50.1 N and longitude113.5 W). A randomized complete block design with three or four replicates of 5 Cms inoculated and 5 uninoculated plants representing 4 to 17 genotypes was assessed each year. The field was not irrigated, fertilization was performed in accordance with locally recommended practices, insecticides were used only when required, and no herbicides were applied. A 2 year potato-cereal rotation was used and all plant debris was buried immediately below the soil surface following harvest. A similar field test was performed at the relatively high humidity isolated plots in 2006 at Bangor, ME to examine results at another geographical location.

Data Acquisition and Analysis

The incidence and severity of foliar disease symptoms were recorded approximately every 2 weeks following planting for Cms inoculated and uninoculated plants as previously described (Kawchuk et al. 1998). The varieties Norland (symptom expressing) and Teton (symptomless carrier) were included each year as a control, but the Teton ratings were excluded from the data set for the NN modeling. Individual plants were rated 0 to 5 where 0 = no visible symptoms, 1 = wilt on lower leaves, 2 = wilt and chlorosis on lower leaves, 3 = wilt on upper leaves, 4 = wilt and chlorosis on upper leaves, and 5 = complete necrosis. Comparison of the randomized inoculated and uninoculated plants confirmed observed symptoms were produced by Cms and the pathogen was also isolated from randomly selected plants each year to confirm the cause of the observed symptoms.

Precipitation and temperature data were obtained from the Environment Canada website http://www.climate.weatheroffice.gc.ca for Claresholm, AB and the Weather Source http://weather-source.com for Bangor, ME. The following inputs were assembled for NN modeling: year, cultivar, expected cultivar response, seeding date, harvest date, plus 40 weather inputs. Expected cultivar response was determined from the average of 20 to 260 field ratings. Weather was tabulated in 2-wk periods, from May 1 to Sept 30 (10 periods) with four weather variables calculated per period, namely: average maximum temperature, C (AvgMaxT), average minimum temperature, C (AvgMinT), average daily temperature, C (AvgT), and total precipitation, mm (TotPcp), to give 40 weather inputs. The predicted output was the visual severity rating (on a 0–5 scale) for BRR on the potato foliage. Model validation was performed with BRR disease data obtained from the ME site for 2006 and the AB site for 2007, 2008, and 2009. Results were examined to determine the ability of the NN model to forecast BRR symptom expression.

Data were arranged in Excel® so that values for the 45 inputs were associated with the severity rating (output) for each plot in each of the years. Each line of 45 inputs paired with a severity rating output formed an example case for training the NN. The total data (example cases) n = 118 was derived from 2,360 plants over 11 years (1986–1996) from the AB site. The data were fairly evenly distributed over the 0–5 severity rating scale (5, 17, 36, 29, 31 cases for the 0–1, 1–2, 2–3, 3–4, 4–5 ratings, respectively). Data were prepared for NN modeling by reducing all input numeric data to one decimal, then randomizing the order of the training data.

NN Modeling

General descriptions of NN modeling can be found in Lawrence (1993), Smith (1993) and Kohzadi et al. (1995). We constructed feed-forward NN models using NeuralWorks Predict® v3.21 software (NeuralWare®, Carnegie, PA) on a desktop PC (2.4 GHz). Briefly, this involved: 1) selecting training and validation subsets; 2) analyzing and transforming data; 3) selecting input and output variables; 4) network construction and training; and 5) model verification. Details on how Predict® conducts NN modeling has been previously published (Hill et al. 2002).

A 10x cross-validation procedure was used to evaluate the NN. That is, 10% of the data were held aside during NN training as a validation set to evaluate the developed models, then this 10% data was returned to the training set and the process repeated by setting aside the next 10% of the data, etc., until every datum appeared once and only once in a validation set. NN inputs and parameter settings were evaluated by averaging net performance across all 10 validation sets.

Neural network modeling is an iterative trial and error process initiated using different random seed numbers. For all predictions, we ran 30 different seed numbers at a time. Usually, no one absolute best NN model was obtained but rather a few different models with similar performance. The ‘best’ NN models were chosen on the basis of a suitable NN architecture (i.e., a minimum number of input neurons connected to a hidden layer that had fewer neurons than the input layer) combined with a high R 2 and a low RMSE when tested against the validation set.

To estimate the relative importance of the different inputs to model predictions, sensitivity analysis was conducted. This Predict® software option allows the determination of the effect that a small change in an input value will have on the output and to numerically rank the inputs according to this sensitivity. Mathematically, the output of sensitivity analysis is a matrix of partial derivatives of output values with respect to input variables. The sensitivity rankings were combined across the 10 ‘best’ cross-validation models. The lowest ranked inputs were manually dropped and the models reconstructed to confirm the lack of influence of the deleted inputs. This process (dropping inputs) was repeated until model performance for predicting the severity ratings dropped off.

Results

Typical BRR symptoms such as wilting and chlorosis of the leaves, stunting of the plants, and breakdown of the vascular tissue in the tuber were observed each year (Fig. 1). Disease symptoms ranging from mild to severe were genotype dependent and highly repeatable in replicated randomized plots in any given year. Symptoms produced by Cms were easily identified and distinguished as they were not observed in the randomized uninoculated controls. Neural network analysis of the data showed that symptom severity was influenced by temperature, precipitation, and genotype susceptibility.

Fig. 1
figure 1

a Early foliage disease symptom development in Clavibacter michiganensis subsp. sepedonicus inoculated potato plants of the cultivar Norland exhibiting wilting and chlorosis of the lower leaves and stunting of the plants. Symptom appearance varied among varieties but the ranking of varieties as to severity of symptoms was consistent. b Disease symptoms in the tubers were relatively infrequent producing a breakdown of the vascular tissues and sometimes cracking of the periderm

Full size image

Determination of Key Environmental Variables

The original 45 inputs could be reduced to just 4 inputs to predict BRR foliar symptom severity and still good NN models were obtained. Preliminary modeling determined that the inputs year, cultivar, and harvest date did not contribute to model solutions. The 10 ‘best’ cross-validation models using the remaining 42 inputs had an average R2 for predicted versus actual severities of 0.94 with SD = 0.04 (0–5 scale). Reduction down to 9 inputs (Table 1) produced an average R2 of 0.95 with SD = 0.04. Further reduction down to 6 inputs (AvgMaxT_Aug1_15, AvgMaxT_Aug16_31, AvgMinT_Jun16_30, TotPcp_Jun16_30, TotPcp_Jul1_15, and cultivar susceptibility) maintained model performance (average R2 of 0.95, SD = 0.04). Models with 4 inputs (AvgMaxT_Aug16_31, TotPcp_Jun16_30, TotPcp_Jul1_15, and cultivar susceptibility) exhibited a slight decrease in model performance (average R2 of 0.93, SD = 0.06) at which point no further inputs were dropped (Fig. 2). The 4-input model was chosen for deployment based on ease of use combined with good performance (Fig. 3).

Table 1 Sensitivity analysis of the parameters in the various models examined. Relative ranking as determined by Predict® sensitivity analysis option
Full size table
Fig. 2
figure 2

Coefficient of determination for predicted with actual severity values for Clavibacter michiganensis subsp. sepedonicus disease symptoms in each cultivated genotype during the years 1986 through 1996. The 4 input NN model was applied and data points shown for each genotype. In general, TotPcpJul1–15 was negatively associated with the expression of BRR disease symptoms whereas TotPcpJun16–30, AvgMaxTmpAug16–31 and Varietal Susceptibility were positively associated

Full size image
Fig. 3
figure 3

Predicted versus actual expression of Clavibacter michiganensis subsp. sepedonicus disease symptoms by year (diamonds = actual rating and squares = predicted values). The field ratings for the symptom expressing Norland control are highlighted with a connecting line. A wide range of reproducible symptoms were observed for the cultivated potato genotypes planted each year and used to develop the NN model (SE = 0 to 0.8). Data collected between 1986 and 1996 was used to develop the deployed NN model. Environmental conditions that produced few disease symptoms (such as those in 1992) reduced disease symptom expression in all lines. Years with lower disease symptom expression may require additional sampling and laboratory testing, especially if the potato cultivar typically produces fewer symptoms. Even lines that typically produce clear disease symptoms may need to be sampled and tested more intensively in those years with environmental conditions producing few disease symptoms

Full size image

To obtain a 4-input deployed model that encompassed the complete range of data, the NN were shown all 108 records during training and the ‘best’ model selected from the 30 NN runs (generated with 30 different seed numbers). This deployed model had an architecture of 6-3-1 (two transformations used for AvgMaxT_Aug16_31 and cultivar susceptibility, one transformation of TotPcp_Jun16_30 and TotPcp_Jul1_15) with a R2 of 0.94 and an absolute average error of 0.2 (0–5 scale). The model was exported from the Predict® software in Visual Basic®, an input/output user-interface added (Fig. 4), and the program compiled into an executable file to facilitate predictions for unknown cases of BRR severity. The user-interface permits the entry of different weather and susceptibility scenarios and the exported NN model runs seamlessly in the background once the user requests the calculation. Although the model was developed for disease symptoms expression in potato foliage, an association was often observed with the symptom expression in the tubers of many cultivated genotypes (Kawchuk et al. 1998).

Fig. 4
figure 4

An image of the executable neural network interface for the input of moisture (TotPcpJun16_30 and TotPcpJul1_15), temperature (AvgMaxTmpAug16_31), and genotype susceptibility (Varietal Susceptibility Rating) variables. Input values accepted are limited to values within the range of variables used in developing the model. The predicted severity output will be between 0.1 with no visible symptoms and 4.8 exhibiting almost complete necrosis. A copy of the executable file is available online at ftp://ftp.agr.gc.ca/pub/outgoing/rb-bh/. UserID is ‘anonymous’ and the password is ‘your_email_address’

Full size image

Relative Importance of Contributing Environmental Factors to BRR Foliar Symptom Expression

Sensitivity analysis was conducted to rank the 4 inputs for their importance to the NN models (Table 1). The AvgMaxT during the last 2 weeks of August was the most important input followed by the TotPcp over the last 2 weeks of June, then TotPcp over the first 2 weeks of July. In general, TotPcpJul1-15 was negatively associated with the expression of BRR disease symptoms whereas TotPcpJun16–30, AvgMaxTmpAug16–31 and Varietal Susceptibility were positively associated. It is interesting that average varietal symptom expression was of lesser importance in our NN models than the three weather inputs (Table 1).

Prediction of Foliar Symptom Expression

Deployment of our best NN model as a transferable executable file should help facilitate predictions of BRR symptom expression in the foliage and the effectiveness of the visual field inspections in any given year. However, for accurate predictions, the range of future input values should be within the input ranges used in our training set (ranges indicated on user-interface, Fig. 4). We successfully applied the model over 4 years and two geographical locations to make BRR severity predictions on pre-elite virus-free seed for 2006 at Old Town, ME and 2007, 2008 and 2009 at Stavely, AB. The average severity predictions generated by the NN for the four site years were 2.0, 2.4, 2.9, and 2.8 with an average absolute difference of 0.15 when compared with the actual average observed ratings of 2.2, 2.4, 2.8, and 2.5, respectively.

Discussion

A major goal of Cms research has been to advance our understanding of disease epidemiology and develop predictive models for symptom expression based on environmental parameters. Previous studies have examined the effect of several individual parameters including inoculum, cultivar, and environment and have provided valuable information identifying factors and interactions influencing the magnitude and onset of disease symptoms. Unfortunately the limited scale of these field trials, in part due to the restrictions imposed in handling a regulated pathogen, has not permitted the development of predictive models for disease symptoms (Bishop and Slack 1987; Westra et al. 1994; Westra and Slack 1994). Our evaluation of a wide range of cultivated genotypes over an extended number of years enabled the examination of an unprecedented number of co-related environmental variables influencing BRR symptom expression. As expected, a predictive model for BRR symptom expression required the application of neural network models over several years of data to form a complex, non-linear system with many co-related environmental variables.

Proportional hazards models have shown that initial symptoms and maximum disease incidence were dependent on cultivar, location, and cultivar by location interactions but not due to increased inoculum (Westra et al. 1994). It was noted that environmental conditions appeared to influence the onset and maximum disease incidence and may have accounted for the variation explained by location. The importance of strategic visual inspections is highlighted by previous studies that showed inspections based on individual parameters such as planting dates were inadequate (Westra et al. 1994). For example, BRR inspections may occur too early to detect disease symptoms or symptoms may be detectable only for a short time due to cultivar senescence that masks disease symptoms. Based on our results, there would be a need to increase sampling and postharvest testing to detect the Cms pathogen in those years with environmental factors producing fewer disease symptoms.

Late season temperature was the most important factor contributing to the severity of the expression of bacterial symptoms in all models (Table 1). Sensitivity analysis of the 4 variable model showed that the AvgMaxT during the last 2 weeks of August was the most influential input followed by the TotPcp over the last 2 weeks of June, then TotPcp over the first 2 weeks of July. This supports previous observations and our understanding of conditions that influence the severity of symptoms of BRR infection (Manzer et al. 1987). Warmer temperatures especially later in the season have been reported to increase disease symptom expression probably due to the impact of restricted water availability caused by the destruction or blockage of vascular xylem tissues by bacteria such as Cms in infected plants (Bentley et al. 2008; De Boer and Slack 1984; Inglis et al. 2000; Laine et al. 2000; Westra and Slack 1992). Extracellular polysacharides and macerating enzymes such as cellulase capable of contributing to the disruption observed plant cell vascular tissues have been identified in Cms. Late season maximum temperatures coincide with increased water transpiration that when combined with vascular destruction would result in elevated disease symptom expression.

Moisture has been shown to influence the expression of bacterial ring symptoms (Dykstra 1941). For example, irrigation has been shown to cause BRR symptoms to develop more rapidly. Results of the present study indicate precipitation, as with irrigation, is an important factor especially earlier in the season. Moisture is critical to bacteria for establishment of infection in plants since propagation, motility, and even structure may be influenced by the hydrodynamics of water (Danhorn and Fuqua 2007). Water is a major factor contributing to propagation and infection by potato brown rot caused by xylem colonizing bacterium such as Ralstonia solanacearum (van Elsas et al. 2001). Moisture would facilitate plant growth, reduce early senescence that may mask disease symptoms, and increase movement of Cms within the infected tissues.

Average genotype symptom expression was ranked of lesser importance in the NN models than the three weather inputs (Table 1). Interestingly, a similar response to environmental conditions was observed by susceptible and resistant genotypes (Fig. 3). Previous studies (Bonde and Covell 1950; Paquin and Genereux 1976; Westra and Slack 1994) also showed that environment is important in the expression of disease symptoms. Our results indicate that foliar symptom expression remained relatively constant among cultivars between years. Location has been shown to modulate inoculum dose response but was not as consistent as the cultivar effect (Westra and Slack 1994). Results indicated that the environmental conditions specific to a location and season influenced BRR symptoms. This location effect suggested environmental conditions influenced symptom expression and provides one explanation for the unexpected appearance of BRR in areas previously considered disease free.

Development of a predictive model for BRR foliar disease symptoms provides an additional tool for integrated pest management to improve detection efficiency by adjusting visual inspection parameters such as timing and frequency. Application of the model will assist regulatory agencies and industry determine the levels of field inspection and postharvest testing required given the environmental conditions and predicted BRR symptom severity. More intensive preharvest and postharvest testing is already recommended in varieties known as symptomless carriers of Cms such as Alpha, Sangre, and Teton. The model also provides further information regarding the environmental impact on the epidemiology of Cms and the relative importance of factors such as moisture and temperature in the infection of potato. Further refinement of the model should be possible with the inclusion of additional environmental data and other factors influencing disease symptom expression.